Research map / Large language models / Language, safety and interpretability
Detecting and Filtering Unsafe Training Data via Data Attribution with Denoised Representation
Yijun Pan, Taiwei Shi, Jieyu Zhao, Jiaqi W. Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DRA denoising to GradSim | GradSim (Llama-3-8B-Instruct) · XSTest-Response detection | AUPRC (%): 1.7 -> 81.5 | Improved |
| apply DRA denoising to RepSim | RepSim (Gemma-2-9B-it) · ToxicChat detection | AUPRC (%): 51.6 -> 98.2 | Improved |
| apply DRA to GradSim before filtering and retraining | GradSim (Gemma-2-9B-it) · XSTest-Response retrain ASR | ASR: 77.0% -> 3.0% | Improved |
| apply DRA to GradSim before filtering and retraining | GradSim (Llama-3-8B-Instruct) · XSTest-Response retrain ASR | ASR: 68.0% -> 43.0% | Improved |
| apply DRA to RepSim before filtering and retraining | RepSim (Llama-3-8B-Instruct) · XSTest-Response retrain ASR | ASR: 63.0% -> 60.0% | Improved |
| apply DRA to RepSim before filtering and retraining | RepSim (Gemma-2-9B-it) · ToxicChat retrain ASR | ASR: 29.0% -> 30.7% | Hurt |
| apply DRA to GradSim | GradSim (Llama-3-8B-Instruct) · Bias in Bios detection | AUPRC: 0.417 -> 0.550 | Improved |
| apply DRA to RepSim | RepSim (Llama-3-8B-Instruct) · Bias in Bios detection | AUPRC: 0.333 -> 0.748 | Improved |
| apply DRA to GradSim before filtering top 500 and retraining | GradSim (Llama-3-8B-Instruct) · Bias in Bios retrain | TPR gender gap: 0.650 -> 0.246 | Improved |
| apply DRA to RepSim before filtering top 500 and retraining | RepSim (Gemma-2-9B-it) · Bias in Bios retrain | TPR gender gap: 0.536 -> 0.346 | Improved |
| use Wildguard moderation classifier for filtering | Wildguard · Bias in Bios | AUPRC: 0.059 | No clear effect |
| increase number of target examples for DRA (ablation) | DRA · jailbreak detection | AUPRC | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Wildguard — WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (NeurIPS 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning and Forgetting Unsafe Examples in Large Language Models (ICML 2024 · similarity 0.88)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.87)
- Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails (NeurIPS 2026 · similarity 0.87)
- Safety Pretraining: Toward the Next Generation of Safe AI (NeurIPS 2025 · similarity 0.87)
- Local Sparsity Enables Unsupervised LLM Safety Detection (NeurIPS 2026 · similarity 0.86)
- HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
